Japanese researchers release 1,000-hour EEG-EMG-audio speech dataset
1020 hours of synchronized brain signals, facial muscle data, and audio from three speakers.
A team of Japanese researchers led by Motoshige Sato has published the largest open dataset of its kind: 1,020 hours of synchronized electroencephalography (EEG), facial electromyography (EMG), and speech audio from three healthy native Japanese speakers. Recordings were captured using three different EEG systems — an ultra-high-density system and two cap-type systems (including eegosports) — spanning 62 to 128 channels across many sessions over several months. Each session provides time-aligned signals, speech-event annotations, and full transcriptions. Technical validation confirmed expected spectral profiles (1/f noise), task-related alpha attenuation, and time-locked evoked responses, ensuring data quality for downstream research.
This dataset is released in the Brain Imaging Data Structure (BIDS) format on OpenNeuro under a CC0 waiver, making it freely available for any use. While primarily motivated by speech decoding applications, the resource also enables work on multimodal signal processing, artifact modeling, cross-device adaptation, and EEG representation learning. For AI researchers, the combination of neural and muscular signals alongside audio offers a rare opportunity to train models that map brain activity directly to speech — a critical step for non-invasive brain-computer interfaces. The open-access nature lowers barriers for reproducibility and accelerates progress in neural speech prosthetics.
- 1,020 hours of synchronized EEG, EMG, and audio from three speakers
- Recorded using three EEG systems (ultra-high-density and cap-type, 62–128 channels) across months
- Released under CC0 on OpenNeuro in BIDS format for speech decoding and multimodal research
Why It Matters
Enables robust speech decoding and multimodal AI training with open-access brain and speech data.